Event Status: Cancelled | Robust and Fuzzy Methods for High-Dimensional Time Series Clustering and Forecasting

Overview

Clustering high-dimensional multivariate time series (MTS) is challenging because variables interact over time, dimensionality amplifies noise, and real data often contain outliers and gradual regime changes that blur cluster boundaries. This talk develops a progression of robust and fuzzy methods for MTS clustering, and shows how the resulting cluster structure can be used to improve forecasting.

We first introduce ROBCPCA, a robust hard-clustering method that extends CPCA to time series by incorporating multi-lag cross-covariances, robust initialization, and a data-adaptive validity index. We then present FCPCA, a fuzzy CPCA-based approach that learns membership-weighted common axes and soft cluster memberships, enabling the capture of overlapping temporal patterns. To handle outliers explicitly, we further develop RFCPCA, a robust fuzzy subspace-clustering family with three complementary strategies: a robust metric, a noise cluster, and trimming.

Finally, we connect clustering to forecasting through a forecastability-aware framework that groups series according to predictive behavior using a robust validation forecasting loss, and then trains cluster-specific deep forecasters. On real benchmark datasets, this yields consistent multi-horizon improvements over both global and individual forecasting baselines for point and probabilistic forecasts.